Ensemble forecasting with IBM Granite time-series models

IBM’s Granite Timeseries Ensemble introduces a new forecasting recipe that merges four complementary Granite time-series models into a single probabilistic forecast. The approach eliminates the need to manually select a single model, which often performs inconsistently across different data types such as industrial signals or retail demand. By pooling quantile forecasts from each member, the ensemble produces median and range predictions without retraining or fitted weights, simplifying deployment for developers.

The ensemble recipe supports two main aggregation methods: the Probabilistic Ensemble, which gives equal influence to each member, and the IQR Ensemble, which weights members by their forecast spread. Both methods operate without historical error fitting, relying instead on fixed rules to combine quantiles. The interface also allows for temperature adjustments and weight caps, enabling experimentation while maintaining a consistent output format for downstream applications.

Evaluation on the GIFT-Eval dataset, using scores captured on 20 October 2026, shows both ensemble recipes outperforming all four individual Granite models on aggregate MASE and average MASE Rank. The Probabilistic Ensemble achieved the lowest MASE (0.683) among the six forecasts compared, while the IQR Ensemble ranked highest in average CRPS Rank. The results highlight the ensemble’s robustness across diverse configurations without requiring prior knowledge of the best-performing member.

The ensemble is implemented through a common Forecaster interface that accepts compatible external models, allowing users to customize member composition or aggregation methods without retraining. The repository also includes additional recipes featuring IBM Research checkpoints under separate license terms. The ensemble’s default configuration and member checkpoints are permissively licensed, with runtime downloads and full details available in the model card.